Fix issue in mul_mat_id for OpenVINO backend

[zhaixuejun1993]

This pull request refines the handling and shape management of expert weights and activations in the OpenVINO backend for the MUL_MAT_ID operation. The core improvements ensure that multi-expert weight tensors retain their full dimensionality, and that input/output tensor shapes are rebuilt more robustly, improving compatibility with dynamic or reshaped input graphs.

Key changes include:

Shape Handling and Tensor Materialization:

• In ggml-decoder.cpp, when materializing non-quantized expert weights, the code now preserves the full reversed 4D shape, ensuring that the expert dimension is not collapsed. This prevents issues where later operations (like Gather/MatMul) would only see a single expert slice.

Dynamic Shape Reconstruction and Reshaping:

• In mul_mat_id.cpp, the logic for squeezing singleton axes from weights, activations, and ids is replaced with explicit dynamic shape reconstruction using ShapeOf and Reshape. This makes the code robust to input tensors that may have undergone reshaping or view operations, ensuring correct logical ranks regardless of input shape permutations.
• The output shape of the MatMul result is now explicitly constructed using dynamic shape information and the expected output rank, replacing previous fixed unsqueeze/squeeze logic. This ensures that the output tensor always matches the required 4D shape, with checks for static rank and row dimension.

General Improvements:

• Added missing include for concat.hpp to support new shape concatenation logic.## Overview

Additional information

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: